HUTECH University

For large asteroid swarms whose orbits must be reconstructed across observation gaps, a graph neural ODE was built with symplectic structure inside the network: instead of approximating conservation laws with soft penalty terms, a Hamiltonian generates the equations of motion directly, and planetary sphere-of-influence radii modulate message passing between bodies. On a one-million-body simulation spanning the Main Belt, Jupiter Trojans, near-Earth asteroids and Centaurs, mean position error is 1.8005×10⁻⁴ AU with energy drift of 8.42×10⁻⁵; on NASA JPL Horizons ephemerides for 99942 Apophis, mean position error across a 60-day blackout is 1.7983×10⁻⁴ AU.

For large asteroid swarms whose orbits must be reconstructed across observation gaps, a graph neural ODE was built with symplectic structure inside the network: instead of approximating conservation laws with soft penalty terms, a Hamiltonian generates the equations of motion directly, and planetary sphere-of-influence radii modulate message passing between bodies. On a one-million-body simulation spanning the Main Belt, Jupiter Trojans, near-Earth asteroids and Centaurs, mean position error is 1.8005×10⁻⁴ AU with energy drift of 8.42×10⁻⁵; on NASA JPL Horizons ephemerides for 99942 Apophis, mean position error across a 60-day blackout is 1.7983×10⁻⁴ AU.